A baseline takes time
A within-subject reference distribution has to be accumulated. In the first hours of a new subject there is no baseline to compare against, and the system has to say so rather than fall back to a population range and pretend.
The framework
Equity by construction
Fairness in a measuring instrument cannot be an intention. It has to be a property of the instrument — something that holds because of how the thing is built, not because of how carefully it is used. These are the four principles Intuitive Pathos is constructed from, and what each one rules out.
Principle 01
Deviation is not deviance.
A measurement that sits far from a population mean is a statement about a distribution. It says: relative to the people who were measured to build this range, this value is unusual. That is a statistical fact and it is often useful.
It is not a statement about the person. Reading it as one — treating distance-from-average as evidence of something being wrong — is a category error, and it is an error that measuring instruments make on our behalf, silently, at scale.
The consequences are not evenly distributed. Whoever was least represented in the cohort that built the range absorbs the most error. Sometimes that is a demographic group. Sometimes it is anyone whose physiology is simply idiosyncratic. The instrument cannot tell the difference and does not try to.
This is not a hypothetical. Race-correction and reference-range bias, and the harm they cause, are documented in the peer-reviewed literature — see the disparity case on the evidence page.
So the first constraint is negative, and it is absolute: in our systems, distance from a population mean is never on its own sufficient grounds to raise a concern about a subject.
What this rules out
- Alerting on a population percentile.
- Cohort-relative scoring as a primary signal.
- Any output whose meaning depends on how unlike other people the subject is.
Canonical statement
The founding statement of this principle is maintained as a controlled document in the regulated flagship’s regulatory file, alongside the mental-health framework it grounds. It is a design input, not marketing copy.
Principle 02
Your reference range is you.
If comparison to a population cannot carry the weight, something else must. The replacement is the per-subject baseline: a within-subject reference distribution built from the subject’s own physiology over time — across rest, activity, recovery, sleep, and whatever else their life contains.
Once that distribution exists, the question the system asks changes completely. Not “how unusual is this person?” but “how unusual is this, for this person?” The first question has an answer that depends on who else was measured. The second does not.
This is also what makes early detection possible at all. A meaningful physiological change often stays comfortably inside a population range for a long time before it crosses any threshold. Against the subject’s own baseline, the same change is visible immediately, because the comparison set is tight.
The baseline is per subject, and it stays with the subject. It is not pooled into a cohort, and it is not a proxy for demographics — the whole point is that it makes demographic proxies unnecessary.
Principle 03
Yang and Yin: symmetric by construction.
Departures from baseline have a direction. Something can be elevated relative to your normal — faster, higher, more — or suppressed relative to it: slower, lower, less. We name these Yang and , and we require the system to treat them as equal and opposite.
That requirement exists because real monitoring systems are rarely symmetric. They are tuned, over years, by what people noticed and responded to — and what people notice is the loud direction. Agitation is instrumented better than withdrawal. Tachycardia better than a subtle downward drift. Escalation better than shutdown.
The asymmetry compounds into a fairness problem, because the quiet failure mode is disproportionately the one experienced by people who are already less likely to be attended to — the sedated, the non-verbal, the exhausted, the very young, the very old, and anyone who has learned that making noise does not help them.
This is a documented clinical pattern, not a turn of phrase — see the hypoactive-delirium citation in the disparity case.
| Loud direction | Quiet direction |
|---|---|
| Agitation, restlessness | Withdrawal, flattening |
| Elevated rate | Sustained downward drift |
| Escalating behavior | Progressive shutdown |
| Acute spike | Loss of normal variability |
In a symmetric model, a given magnitude of departure from the subject’s own baseline carries identical weight in either column. Direction is preserved in the output — it is clinically and behaviorally meaningful — but it never scales the severity.
Principle 04
Architectural impossibility.
The strongest guarantee a system can offer is not that it will not do something. It is that it cannot.
Most safety properties in software are promises: a policy, a review step, a configuration flag, a line in a manual. Promises degrade. They are waived under pressure, misconfigured during integration, and forgotten when the person who understood them moves on.
Some properties are too important to leave in that category. For those, we make the harmful state unrepresentable — encoded in the type system and the module boundaries, so that violating it is not a policy breach but a program that does not compile.
This is a deliberately expensive choice. It constrains what we can build, and it makes some features simply unavailable. That is the point: a constraint you can lift on a bad day was never a constraint.
Made structurally impossible
- Cohort comparison as a signal path. There is no interface through which a demographic cohort can become a subject’s reference distribution.
- Identifiable physiology crossing the federation boundary. What federates and what never federates is a typed contract, not a setting.
- Clinical conclusions from non-clinical products. The wellness and veterinary products have no access to the regulated decision logic; it is not in their dependency graph.
- Asymmetric severity scaling. The interfaces do not admit a direction-weighted severity term.
Fairness you can switch off is not fairness. It is a setting.
Which is why the guarantees on this page are enforced in the type system rather than promised in a policy — and why some features we could build simply do not exist.
The mathematics
Regime change, not threshold crossing.
Asking “has a number crossed a line?” is a late question. Asking “is this system changing state?” is an earlier one, and it is the question the Wu primitives are built to answer. It rests on established science — see the evidence.
Where the mathematics lives
These primitives have exactly one implementation, in an internal library
(wu-oracle) that is the single source of truth for the
closed-form numerics. Every product reaches them through typed wrappers
in the shared substrate rather than reimplementing them — which
means a correction to the mathematics is made once and propagates by
version bump. See the platform architecture.
Limits
What this framework does not solve.
A claim about fairness that comes with no limits is a marketing claim. Here are ours.
A baseline can be built on a bad period
If a subject’s observation window is itself abnormal, their baseline encodes that. Detecting and handling this is a real, unsolved-in-general problem, not a footnote.
Sensors carry their own bias
Per-subject comparison removes the population-reference problem. It does not remove measurement bias in the sensor itself — a device that measures some people less accurately still does so. See the evidence.
Fair input is not fair outcome
An equitable measurement feeds into a care system with its own inequities. We are addressing one link in that chain and we do not claim to be addressing the chain.
Questions
Common questions
Isn’t comparing someone to themselves just a moving average?
No. A moving average smooths a signal; it still needs a threshold to mean anything, and that threshold usually comes from a population. A per-subject baseline is a full within-subject reference distribution — it supplies the comparison set itself, which is what the population range was previously doing.
Doesn’t ignoring population data throw away useful information?
Population data remains useful for many things — epidemiology, resource planning, deciding what is worth measuring at all. The constraint is narrower than it sounds: population distance may not, on its own, be the basis for a conclusion about an individual subject. That is a statement about inference, not about data.
How is this different from “personalized medicine”?
Personalization usually means selecting a sub-cohort more precisely — still comparing you to a group, just a smaller one. This framework removes the group from the comparison entirely. Your reference distribution has one member.
Why the Yang/Yin naming?
Because the property being named is symmetry between two directions that are equal and opposite, and the pair carries that meaning without importing a value judgment. “High and low” and “positive and negative” both smuggle in a better and a worse. The whole point is that neither direction is the bad one.
Is any of this cleared for clinical use today?
No. The clinical system, Pathos Vigil, is investigational and has not been cleared or approved by the FDA or any other regulatory authority. The non-clinical products make no diagnostic or therapeutic claim by design. We would rather state the stage plainly than imply otherwise. See the clinical overview for the current regulatory posture.